Cost
An unsupported claim often needs another turn, or a person, to check it before you can use the reply.
Map · Foundations · 08 · Hallucination
Concept card · Foundations
Hallucination
A hallucination is a reply that states something the prompt and the sources do not support.
A reply that states something the prompt and the sources do not support.
What changed
Oct 17: First card. The picture marks a claim that is in the source and a claim that is not. Sources are Anthropic's reduce-hallucinations guide, Gemini's safety guidance, and OpenAI's accuracy guide.
1 What it is
The model can write a fluent sentence that was not in the prompt and not in the source you attached. That unsupported claim is a hallucination. A claim that contradicts the source is one too.
2 Why it matters
Quality
A fluent sentence is not a source. The claim is only as good as the text that supports it.
Risk
If the product acts on the reply, an unsupported claim can become a wrong decision. Checks reduce that chance. They do not remove it.
3 How to apply · try today
Take one reply and list each factual claim. For each claim, point to a sentence in the prompt or the source. If you cannot, mark the claim unsupported.
Common mistake
Treating a confident or specific sentence as proof that the claim was in the source.
Self-check
The source says the warehouse closes at 18:00 and does not mention Sunday. The reply says it is open on Sunday. What is the Sunday claim?
Coming soon. One reply you already use, with each claim marked supported or not.
Coming soon. How a provider's grounding or citation tool relates to checking a claim yourself.
Questions
- What is a hallucination?
- A hallucination is a reply that states something the prompt and the sources do not support.
- Does fluent wording make a claim supported?
- No. Support is a sentence in the prompt or the source, not the tone of the reply.
- Do checks remove hallucinations?
- No. They reduce the chance you act on an unsupported claim. They do not remove it.